The NVIDIA Collective Communication Library (NCCL) is a critical software layer enabling high-performance collectives on large-scale GPU clusters. Despite being open source with a documented API, its internal design remains largely opaque. The orchestration of communication channels, selection of protocols, and handling of memory movement across devices and nodes are not clearly understood, making it difficult to analyze performance or identify bottlenecks. This paper presents a comprehensive analysis of NCCL, focusing on its communication protocol variants (Simple, LL, and LL128), the mechanisms governing intra-node and inter-node data movement, and ring-and tree-based collective communication algorithms. The insights obtained from this study serve as the foundation for ATLAHS, an application-trace-driven network simulation toolchain capable of accurately reproducing NCCL communication patterns in large-scale AI training workloads. By demystifying NCCL's internal architecture, this work provides guidance for system researchers and performance engineers working to optimize or simulate collective communication at scale.

Network simulators play a crucial role in evaluating the performance of large-scale systems. However, existing simulators rely heavily on synthetic microbenchmarks or narrowly focus on specific domains, limiting their ability to provide comprehensive performance insights. In this work, we introduce ATLAHS, a flexible, extensible, and open-source toolchain designed to trace real-world applications and accurately simulate their workloads. ATLAHS leverages the Group Operation Assembly Language (GOAL) format to model communication and computation patterns in AI, HPC, and distributed storage applications. It supports multiple network simulation backends and handles multi-job and multi-tenant scenarios. Through extensive validation, we demonstrate that ATLAHS achieves high accuracy in simulating realistic workloads (consistently less than 5% error), while significantly outperforming AstraSim, the current state-of-the-art AI systems simulator, in terms of both simulation runtime and trace size efficiency. We further illustrate ATLAHS’s utility via detailed case studies, highlighting the impact of congestion control algorithms on the performance of distributed storage systems, as well as the influence of job-placement strategies on application runtimes.

Aiming at the problems of low data rate and limited battery power of traditional underwater wireless sensor network (UWSN) based on acoustic communication, a resource allocation strategy for magnetic induction (MI) communication-based UWSN in ocean current scenario is investigated in this paper. Specifically, multiple sensor nodes (SNs) are distributed in seawater at different depths. First, an autonomous underwater vehicle (AUV) is introduced to charge the SNs by magnetic coupling resonant wireless power transfer (MCR-WPT) technology. Then, the SNs transmit data to the AUV through MI communication. The influence of the ocean current on the SNs and AUV is analyzed. A genetic algorithm-based particle swarm optimization algorithm is used to optimize the AUV navigation trajectory. After collecting the data from the SNs, the AUV moves to the position under a surface base station and sends the collected data to the surface base station by MI communication. To minimize system energy consumption, the transmitting power of AUV and SNs are jointly optimized under the constraints of energy causality and transmitting power. The suboptimal solution to the formulated optimization problem is obtained by adopting the improved sparrow search algorithm (ISSA) combining Cauchy variation and reverse learning. Simulation results show that the ISSA has lower system energy consumption than other benchmark methods.

Sensor nodes (SNs) in underwater wireless sensor network (UWSN) may generate massive amounts of data and have computing task requirements. In this paper, mobile edge computing (MEC) is considered in UWSN. A resource allocation strategy in autonomous underwater vehicle (AUV)-aided MEC system is investigated to minimize the total system energy consumption. Specifically, the MEC servers are deployed at both AUV and buoy floating on the sea. By adopting magnetic induction (MI) communication technology, the AUV collects computing tasks from different SNs. Partial tasks are computed at the AUV, while the rest of tasks are offloaded to the buoy. The trajectory of the AUV, the transmitting power of SNs and the AUV, and the offloading ratio coefficient are jointly optimized. The suboptimal solution to the formulated optimization problem is obtained by using a social learning particle swarm optimization (SL-PSO) algorithm. Simulation results show that the proposed strategy has a faster convergence speed and better global search capability compared with other baseline schemes.

Aiming at the problem of low data rate in autonomous underwater vehicle (AUV)-based underwater acoustic communication system, a hybrid magnetic induction (MI) and reconfigurable intelligent surface (RIS)-assisted communication strategy is proposed to maximize system channel capacity. Specifically, an AUV first collects data from all the seafloor nodes by adopting the MI communication technology. Then, the AUV forwards the data to another AUV carried with a RIS by using the underwater acoustic communication method. With the help of the RIS, a strong reflective path between the first AUV and a surface base station (BS) on the sea is formed. The surface BS could receive the data at a relatively high data rate. In order to maximize the system channel capacity, the acoustic incidence angle, the distance between the first AUV and the RIS, the distance between the RIS and the surface BS, acoustic signal frequency, and transmitting power are jointly optimized. The formulated optimization problem is solved by employing a butterfly optimization algorithm (BOA) and an improved butterfly optimization algorithm (IBOA), respectively. Simulation results show that the IBOA can increase the system channel capacity more effectively than the basic BOA.

The traditional underwater wireless sensor network (UWSN) based on acoustic communication has the shortcomings of low data rate and limited battery power. In this paper, hybrid acoustic and magnetic induction (MI) communication are considered to overcome the above drawbacks. A resource allocation strategy in autonomous underwater vehicle (AUV)-assisted edge computing UWSN is investigated to minimize the total system delay. Specifically, all the sensor nodes (SNs) are divided into different clusters. The SNs within a cluster send the data to the cluster head (CH) via the acoustic communication. The CH forwards the data to the AUV by the MI communication. Then, the AUV moves to the position under a surface vehicle (SV) carried with a edge server. The AUV forwards the data to the edge server through the MI communication. The transmitting power, channel bandwidth, and computational resources are jointly optimized. The formulated non-convex optimization problem is solved by using an alternating iterative optimization algorithm. Compared with other schemes, the proposed strategy can reduce the total system delay more effectively.

Traditional underwater wireless communication in single medium has the limitations of low rate and high delay. In this paper, magnetic induction (MI)-based cross-medium communication is taken into account to reduce the transmission delay. Specifically, multiple autonomous underwater vehicles are used to collect data from underwater sensor nodes by MI communication. The collected data is directly transferred to a unmanned aerial vehicle above the water via ultra-low frequency MI communication. The cross-medium data collection and transmission problem is formulated an optimization problem. The objective is to minimize the total delay under the constraints of transmitting power, transmission distance, and number of turns of MI coil. A standard particle swarm optimization (SPSO) algorithm and a quantum-behaved particle swarm optimization (QPSO) algorithm are adopted to obtain the suboptimal solution, respectively. Simulation results show that the QPSO algorithm is superior to the SPSO algorithm in reducing the total delay.

Given the rapid growth of the mobile internet market, high-priced telecom packages are limiting the growth of mobile internet use. Telecom operators and internet companies are working together to offer free-flow packages. This paper proposes a complete content selection and package pricing scheme (CSPPS) based on data provided by a telecom provider. This paper first conducts a feasibility analysis on the free traffic package scheme. By analyzing users’ usage detailed records (UDRs), this paper identified a model for user internet behavior. A scheme for content selection was developed based on this model. To discover users’ online timing characteristics, we developed a tensor decomposition factor model. To complete the package pricing process, a two-sided market model and SoftMax algorithm were combined. Performance evaluations were conducted using real data from China Unicom. According to our experimental results, our package selection scheme is effective for different scenarios.

In recent years, deep learning has achieved remarkable results in the text-image retrieval task. However, only global image features are considered, and the vital local information is ignored. This results in a failure to match the text well. Considering that object-level image features can help the matching between text and image, this article proposes a text-image retrieval method that fuses salient image feature representation. Fusion of salient features at the object level can improve the understanding of image semantics and thus improve the performance of text-image retrieval. The experimental results show that the method proposed in the paper is comparable to the latest methods, and the recall rate of some retrieval results is better than the current work.

User portrait analysis is one of the key points in human behavior analysis. It is important to describe or guess user's characteristics through rational methods in business analysis. In this paper, we use the user details records data set from a mobile operator to analyze the preference of users with different brand phones for different APPs and propose the concept of mobile Internet life personas (MILP) and the latent MILP indexing (LMILPI) model for the analysis of users' MILP. At the same time, we build user portrait analysis framework based on the latent semantic indexing (LSI) theme model, LMILPI model, and association rule mining. On the one hand, we analyze users' preference for APP content when using different mobile brands. On the other hand, we analyze the relationship among mobile brands, user access time, and MILPs to describe users' Internet behavior. Our research shows that there is a difference between users who use different brands of mobile phones: 1) users who use different brand phones have different preferences for different APPs. However, if mobile brand marketing methods or target users are same or similar, the APP preference of these brands will be similar; 2) MILPs are different between users who use different brands of mobile phones, but the MILPs displayed on Android platform are similar even though brands are not same, while the MILP displayed on iPhone is quite different from MILPs on Android; and 3) MILPs' importance will be changed by mobile phone brands and time periods. The analytical framework which we propose can provide commercial solutions such as application recommendations, market strategy formulation, Internet access, and other fields.